usage: structure init [<name>] [--repo <path>] [--name n] [--describe d] [--yes]
structure init --project [--name n] [--describe d] [--starter s]
structure init --demo [--repo <path>]
structure init --dry-run [--project] [--adopt|--fresh]
Stand up a company: a short interview, a tailored structure, a typed
scaffold and the compiled agent dialects. <name> makes ./<name>/ and
stands it up INSIDE that folder; bare `new` interviews here, and `new .`
is that spelled out. A name and --repo both answer where it goes, so
together they are refused rather than ranked.
Two doors, one command: a repo already holding CLAUDE.md, AGENTS.md or
.cursorrules is offered [a]dopt — canon distilled from your own files,
originals preserved, nothing committed — or [f]resh, a blank typed
scaffold. --adopt / --fresh chooses without being asked.
--project a place to begin instead of a company: five files, the same
substrate and the same gating, a smaller layout. `structure
graduate` adds the rest in place when it stops being just you.
--demo an example company, already at work. Invented people, invented
work, no model calls; refuses --name/--describe/--runner.
--dry-run say what it would do and do none of it: the files it would
create, the files it would rewrite and where each original
would be kept. Writes zero bytes — no scaffold, no git init,
no commit, no folder, and this device does not record the
repo. Asks nothing; every question takes its default.
--starter form the company (or project) from a shipped starter: its
spine, functions, roles and skills arrive as copies, each
carrying one provenance line naming the starter. One code
path, both doors — they differ only in gating and the four
project files.
--yes take every default and ask nothing (non-interactive).
--runner claude-cli|codex-cli|ollama — which runner drives the
model-first arc: the follow-up conversation, the selection
of functions/roles/skills/document types, and the canon
drafts (in parallel). With no runner the surface says
`No runner · starter defaults` and template seeds stand.
--no-smart heuristics only; nothing is asked of a model.